__author__ = 'mcapizzi'

import numpy as np

import tensorflow as tf

#placeholders and run

#can set default dimensions for placeholder, or leave unnamed

spot1 = tf.placeholder("float64")

spot2 = tf.placeholder("float64")

spot3 = tf.placeholder("float64", shape=[None, None]) #don't need to type, but helps safeguard

add_op = tf.add(spot1, spot2)

add_op2 = tf.add(spot1, spot3)

sess= tf.Session()

output = sess.run(add_op, feed_dict={spot1: 1., spot2: [.1, .1]})

#if explicitly type, then need to make sure dimensions match

output2 = sess.run(add_op2, feed_dict={spot1: 1., spot3: np.array([[.1, .1]])})

output3 = sess.run(add_op2, feed_dict={spot1: 1., spot3: [[.1,.1]]})

#variables

#must initialize values

v1 = tf.Variable(np.zeros((5,5)))

init_op = tf.initialize_all_variables()

up = tf.assign(v1, v1 + 5) #updates values of variable

#don't forget to initialize variables!

sess.run(init_op)

for i in range(3):

print v1._AsTensor().eval(session=sess)

out = sess.run(up)

print type(out)

print type(v1)

print v1._AsTensor().eval(session=sess) #allows you to see the value of the variable

#built in neural network functions

#activation

tf.nn.sigmoid()

# tf.nn.tanh()

# tf.nn.relu()

#cost

#tf.nn.sigmoid_cross_entropy_with_logits

#tf.nn.softmax

#tf.nn.softmax_cross_entropy_with_logits

#BUILD FEEDFORWARD WITH NO ACTIVATION

#BUILD COST WITH tf.softmax_cross_entropy_with_logits(FF)

#BUILD PREDICT WITH tf.softmax(FF)

tf.nn.embedding_ops